reconstitute spatial weights matrix...
TRANSCRIPT
FULL PAPER PROCEEDING Multidisciplinary Studies
Available online at www.academicfora.com
Full Paper Proceeding BESSH-2016, Vol. 76- Issue.3, 35-45 ISBN 978-969-670-180-4
© 2016 The Authors. Published by Academic Fora. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under responsibility of the Scientific & Review committee of BESSH-2016.
BESSH-16
RECONSTITUTE SPATIAL WEIGHTS MATRIX CONSIDERING U-CITY
NETWORK STRUCTURE
SUNGHWAN KIM1, KABSUNG KIM2*
1,2,3Yonsei University, South Korea
Keywords: Spatial Weights matrix (SWM)
Spatial Lag Model (SLM)
Housing Sales Price, Seoul
U-City
Abstract. In this study, we reconstitute a spatial weights matrix for U-City Network Structure and construct a
regression-based spatial analysis model among 16 major administrative areas. Existing spatial weight matrices
for spatial autoregressive models commonly consider adjacency or invert distance weighting. However,
nowadays, a virtual network system is more important than actual distance to reveal spatial dependencies.
Especially in Korea, U-City(or smart city) network structures have expanded rapidly by the government and
reconstructed traditional urban systems. The most characteristic point from advanced research is that we
consider the U-City network structure into the spatial adjacencies and interactions model. This result in a family
of spatial OD models that represent an extension of the spatial regression models described in Anselin(1988).
INTRODUCTION
-CITY(UBIQUITOUS CITY), the Korean smart city model, was
made to take advantage of the developed Korea‟s Information
technology(IT) infrastructure and to make urban management
easier such as disaster prevention and crime prevention. U-City
projects in Korea have been operated by government test-bed
projects and are underway to build a nationwide (MOLIT, 2013).
In particular among the projects, CCTV integrated management
service to take care of crime-ridden districts getting responses
from people. The most important reason for attracting attention to
U-City is the concept of „physical distance‟ is weakening.
Because older cities are now changing into IT infrastructure basis
management environment.
Housing sales price, a research subject of this study, can also be
seen evolving change pattern with U-City system. Since housing
fixes on the land, it has a unique feature that cannot move. Thus,
house sales prices have not been determined independently and
usually formed by exchange influences among the neighbored
regions; it is widely accepted theory(Park and Ahn, 2005; Gillen
et al., 2001). However, housing prices so that Park and
Kim(2007) has pointed out, sometimes showing another
correlation with it other than neighboring or physical distance. As
a typical example on Seoul, house sales prices in the Gangnam-gu
affects the Mokdong and Jamsil area nearly without delay. Look
at this on the basis, in addition to the regional neighbor and
* Corresponding author: Kabsung Kim
E-mail: [email protected]
distance, several other influential acts believed to form the house
prices.
Figure 1. An outline map of the research area(Seoul, Korea)
U
Kabsung Kim /BESSH--2016/Full Paper Proceeding Vol No-76, Issue 3, 35-45
International Conference on “Business Economics, Social Science & Humanities” BESSH-2015
36
In this study, to recognize the fact that such changes in the U-City
and another urban model is likely to affect home sales, new
spatial weights matrix (SWM) that can reflect the influence would
like to propose. In addition, we have been studying with an
emphasis on understanding the statistical significance and
efficiency of the new SWM.
THEORETICAL DISCUSSIONS AND PREVIOUS
RESEARCHES
Studies in regards to real estate values have generally taken two
theoretical flows. While one flow follows the equilibrium price
model based on the law of demand and supply, another one
centers around the hedonic price model by Lancaster (1966) and
Rosen (1974).
The hedonic price model argues that the purchase of any good
signifies the purchase of the entire good, including its inherent
characteristics, and thus the price of a good is decided by its
quantity and inherent characteristics. The hedonic price model is
useful for a wide research subject as it can analyze the various
elements affecting real estate prices. Within these logical
boundaries, statistical methods such as regression models were
applied to take into consideration the various characteristic of a
house including generational, regional, and apartment-specific
features. Of the various characteristics, the environmental element
was studied in regards to unwanted public facilities such as trash
incineration facilities (Cho, 1998) or detention centers and
electricity substations (Choi et al, 2000) and their effects on the
real estate prices. Preceding researches that studied real estate
prices caused by social environmental
changes from public rental housings also applied these methods
(Woo, 2005; Moon et al, 2006; Hwang, 2009).
Studies that follow the hedonic price model evolved alongside the
development of improved statistical instruments. The most
notable case is the use of hierarchical linear model. Hedonic price
model fundamentally applies several variables of different levels
including generational, regional, and apartment-specific features.
Naturally, as it provides the most suitable tool for analyzing the
variables, hierarchical lineal model has been frequently used in
recent real estate price researches based on hedonic price model
(Kim, 2003; Kim et al, 2004; Choi et al, 2004; Jung, 2006; Lee et
al., 2012; Yoon et al., 2012). Another recurrently attempted tool
is the spatial regression analysis, which can take spatial
autocorrelation into consideration (Choi et al, 2003; Oh et al,
2014). These attempts are results of the due to the immobility of
the real estate, which necessitates control over the real estate
prices‟ spatial autocorrelation.
On the other side, about an SWM, there is also extensive
literature. When it is first introduced by Anselin (1988), there are
only a few generating methods utilizing adjacent or distance
several generation methods exist. The simplest types of SWM are
contiguity type matrices. However, over time, new generation
methods have been developed. Getis and Aldstadt (2004)
suggested over a dozen different types of SWM. Nowadays, the
geo-statistical method is known as the most complex type. In
between these differences are a host of different distance-related
functions. Matters should focus on the debate on SWM is not a
using complex or advanced method but specifying how research
object is located in space concerning the other reference object.
There are 3 types of considerations and viewpoints when making
SWM (Aldstadt and Getis, 2006):
1. “A theoretical notion of spatial association, such as a
distance decline function.”
2. “A geometric indicator of spatial nearness, such as a
representation of contiguous spatial units.”
3. “Some descriptive expression of the spatial association
within a set of data, such as an empirical variogram
function.”
In consideration of the above three perspectives on generating
SWM, this study focuses on the third viewpoint. Because
viewpoints 3 is more proper way than the others, perhaps. It can
be proved in descriptive analysis and phenomenon-based
researches.
PHENOMENON-BASED RESEARCH
Data Acquisition
The data used in this study can be divided mainly into two types.
At first, the actual transaction data for calculating home sales was
provided by「 The Actual Acquisition Price of Real Estate
Information System」 which operates in MOLIT. In contrast to
previous studies (Park and Kim, 2007; Lee and Park, 2013), the
actual transaction data is used in the present study. Thus, it is
expected to be a more sophisticated analysis can be done by
reflecting the real price, not a call price. Data included three
months from October to December 2015, and the total 19,131
sales data were collected. Data covered Seoul, Korea.
Kabsung Kim /BESSH--2016/Full Paper Proceeding Vol No-76, Issue 3, 35-45
International Conference on “Business Economics, Social Science & Humanities” BESSH-2015
37
TABLE I
THE RESULT OF DESCRIPTIVE ANALYSIS
Variables Measure Mean SD Min. Max
Dependent
Variables
(total 14
variables)
Physical
Characteristics
(total 10
variables)
Total Area ㎡ 112.14 34.09 15.00 350.50
Rebuilding is
Scheduled
Yes(%) 0.01 0.09 - -
Total Buildings 5.32 8.11 1.00 74.00
Area for Exclusive Use ㎡ 78.50 28.48 11.96 270.25
Num. of Rooms 2.88 0.79 1.00 7.00
Num. of Baths 1.52 0.57 1.00 3.00
Total Household 412.82 548.98 45.00 5,540.00
Parking Lot/HH 1.20 0.49 0.30 5.80
Porch Cascade
Type
0.63 0.37 - -
Aging of Apts. Years 12.38 9.18 0.00 36.00
Environmental
Characteristics
(total 4 variables)
Dist. to General
Hospital
Meters 1,732.15 812.33 133.86 4,552.27
Dist. to Primary School Meters 452.31 223.57 78.52 1,243.54
Dist. to High School Meters 720.94 391.08 52.22 3,005.38
Dist. to Subway Station Meters 635.30 262.41 42.29 2,802.51
U-City Available Yes(%) 44.00 0.13 - -
Independent
Variable KRW/㎡ 10thou.
KRW 653.00 289.80 178.00 3,064.00
TABLE II
THE AVERAGE PRICES OF HOUSING SALES PRICES
Administrative Districts Prices (in million KRW)
Gangnam 1010
Seocho 1004
Yongsan 832
Songpa 714
Mapo 627
Gwangjin 606
Jung 564
Seongdong 553
Jongno 551
Yangcheon 528
Dongjak 517
Gangdong 485
Yeongdeungpo 457
Seongbuk 418
Dongdaemun 404
Seodaemun 394
Gangseo 384
Eunpyeong 374
Gwanak 371
Gangbuk 348
Guro 344
Jungnang 315
Geumcheon 315
Nowon 297
Dobong 279
* Price : High to Low
Kabsung Kim /BESSH--2016/Full Paper Proceeding Vol No-76, Issue 3, 35-45
International Conference on “Business Economics, Social Science & Humanities” BESSH-2015
38
Figure 2. Apartment distribution map of the research area
Then, the building register to obtain the co-housing estates
information was constructed in the database. This data was
processed download from the large-capacity data providing
system of "Architectural Information Open System", which
operates in MOLIT. The building that has been registered as a
"co-housing" in the construction in the December 2015 criteria
Seoul, a total 96,603 buildings and this is a value except for the
duplicates.
Next, it was processed and join in the parcel number unit (PNU)
code for the purpose of trying to analyze together the actual
acquisition price information and information of the co-residential
complex to prepare for those mentioned above. Matching rate is
100%, all of the information has been fully joined.
Finally, to calculate the other environmental variables, get the
(X, Y) coordinates of subway stations and bus stops from
geographic information system (GIS) shape file. Make the GIS
spatial operations utilization to analyze the traffic environment,
the educational environment, and the living environment.
Basic statistics of data obtained through these process is as
<Table 1> below. Which is a feature that was the rebuilding
related variable is added that is not described in the previous
variable explanation. According to the Park and Kim (2007),
general buildings prices decline gradually over time because the
life expectancy due is approaching, but the apartments prices are
not when if reconstruction plan exists. Therefore, after the
„reconstruction plan‟ variable has inserted as a control value, then
the analysis was performed.
Graphical Analysis
As a result of comparing the distinction apartment sale price, it
has been derived as <Table 2>. Gangnam shows highest average
sale price in Seoul, and then Seocho, Songpa, and Yongsan form
high price group. Nowon has lowest mean sales price, and then
Dobong, Geumcheon, and Eunpyeong join small price group. The
average sale price difference between Dobong District(showed
the lowest cost, 279 million KRW) and Gangnam-gu, showed the
highest amount of money (1,010 million KRW) was found to be
up to about 3.6 times.
After a visual analysis of time-series change over the study
period, the graphs in the fourth quarter of 2015 showed that draws
a steady graph does not cause noticeable changes. Therefore, it
can be analyzed by the cross-sectional method to derive the mean
quarterly without including the features of the time-series data.
The distinct feature is that buying and selling price has a
downward trend in the Gangseo (Magok district), which is
expected to intensive housing supply in the future (about a year).
For the reason of this trend appeared, Magok district has lots of
long-term periods chonsei housing and public rental housing, so it
has an adverse impact on the price in the vicinity of the apartment
(Kim et al, 2015).
Correlation Test
It was carried out correlation test to explain the similarity of these
distinctions sale price movement statistically. First of all, looking
at the area that showed a high correlation, Seocho-Gangnam-
Sangpa region, as we saw in the previous graphical situation
analysis had been given a strong influence each other. This
influence could be explained by the proximity of the group.
However, when seen in the reference to the Gangnam, it was
found that has signed Yangcheon, Yeongdeungpo, a high
correlation also Nowon. Despite the considerable distance away
from each region, it seems to be having an unknown impact in
determining the sale price fluctuations and some relationships
between each sphere.
Kabsung Kim /BESSH--2016/Full Paper Proceeding Vol No-76, Issue 3, 35-45
International Conference on “Business Economics, Social Science & Humanities” BESSH-2015
39
Figure 3. Correlation map with Gangnam and the others
The Yangcheon District that shows the distinctive look.
Yangcheon, as seen in the following <Figure 4>, in addition to
the Mapo and Yeongdeungpo, which are adjacent to each other
and also appeared receive affecting both Gangnam and Seocho.
This seems to be due to Mokdong Newtown is located in the
Yangcheon. Mokdong is a large-scale apartment complex that has
been construction in 1985-1988, and forms a high level of house
prices in the Yangcheon as education fever is a very high.
Because Gangnam•Seocho shares with this feature and is
estimated that show such a correlation. That it might seem that
the correlation and also due to some other influence, except for
the fact that geographical proximity can see in the example
Yangcheon.
Figure 3. Correlation map with Yangcheon and the other
Margareth Gfrerer /BESSH--2016/Full Paper Proceeding Vol No-76, Issue 3, 1-10
International Conference on “Business Economics, Social Science & Humanities” BESSH-2015
Full result of correlation test is as follows <table 3>.
* GN : Gangnam, SC : Seocho, YS : Yongsan, SP : Songpa, MP : Mapo, GJ : Gwangjin, JU : Jung, SD : Seongdong, JO : Jongno, YC : Yangcheon, DJ : Dongjak GD : Gangdong, YP : Yeongdeungpo, SB : Seongbuk, DM : Dongdaemun, SM :
Seodaemun, GS : Gangseo,
EP : Eunpyeong, GA : Gwanak, GB : Gangbuk, GR : Guro, JN : Jungnang, GC : Geumcheon, NW : Nowon, DB : Dobong
TABLE III THE RESULT OF CORRELATION TEST
THE RESULT OF CORRELATION TEST GN SC YS SP MP GJ JU SD JO YC DJ GD YP SB DM SM GS EP GA GB GR JN GC NW DB
GN 1.00
SC 0.97 1.00
YS 0.66 0.71 1.00
SP 0.95 0.97 0.61 1.00
MP 0.56 0.59 0.93 0.55 1.00
GJ 0.81 0.77 0.61 0.78 0.42 1.00
JU 0.38 0.43 0.92 0.44 0.65 0.51 1.00
SD 0.72 0.68 0.42 0.65 0.48 0.66 0.38 1.00
JO 0.55 0.64 0.73 0.57 0.82 0.36 0.82 0.47 1.00
YC 0.91 0.88 0.69 0.84 0.86 0.75 0.29 0.67 0.42 1.00
DJ 0.86 0.91 0.87 0.81 0.77 0.78 0.29 0.65 0.49 0.91 1.00
GD 0.66 0.65 0.31 0.64 0.51 0.33 0.28 0.45 0.24 0.37 0.39 1.00
YP 0.91 0.89 0.73 0.86 0.89 0.71 0.54 0.70 0.45 0.94 0.86 0.35 1.00
SB 0.58 0.61 0.55 0.55 0.42 0.64 0.73 0.66 0.70 0.81 0.53 0.48 0.67 1.00
DM 0.43 0.39 0.64 0.40 0.43 0.71 0.89 0.63 0.81 0.72 0.60 0.77 0.82 0.97 1.00
SM 0.53 0.54 0.76 0.58 0.88 0.69 0.80 0.56 0.79 0.77 0.54 0.69 0.83 0.61 0.87 1.00
GS 0.30 0.33 0.45 0.33 0.66 0.57 0.41 0.20 0.33 0.56 0.49 0.54 0.69 0.45 0.49 0.41 1.00
EP 0.41 0.48 0.60 0.49 0.42 0.22 0.69 0.40 0.78 0.51 0.45 0.41 0.59 0.59 0.66 0.79 0.71 1.00
GA 0.61 0.62 0.78 0.57 0.37 0.40 0.51 0.66 0.33 0.28 0.71 0.63 0.65 0.45 0.42 0.60 0.77 0.80 1.00
GB 0.59 0.64 0.59 0.68 0.63 0.70 0.66 0.59 0.61 0.69 0.78 0.42 0.63 0.93 0.82 0.69 0.51 0.68 0.44 1.00
GR 0.34 0.37 0.62 0.41 0.55 0.44 0.58 0.72 0.40 0.55 0.51 0.59 0.43 0.34 0.33 0.64 0.78 0.55 0.88 0.26 1.00
JN 0.39 0.36 0.25 0.22 0.21 0.57 0.46 0.80 0.59 0.41 0.33 0.48 0.50 0.66 0.86 0.23 0.39 0.77 0.51 0.69 0.38 1.00
GC 0.28 0.23 0.55 0.42 0.54 0.48 0.43 0.51 0.37 0.35 0.45 0.38 0.51 0.81 0.73 0.56 0.70 0.81 0.89 0.39 0.88 0.69 1.00
NW 0.88 0.80 0.62 0.86 0.59 0.80 0.29 0.77 0.48 0.80 0.85 0.49 0.81 0.88 0.85 0.58 0.66 0.71 0.45 0.93 0.40 0.74 0.31 1.00
DB 0.58 0.56 0.41 0.51 0.37 0.61 0.51 0.86 0.68 0.55 0.48 0.34 0.42 0.70 0.59 0.31 0.25 0.76 0.43 0.69 0.54 0.88 0.52 0.64 1.00
Margareth Gfrerer /BESSH--2016/Full Paper Proceeding Vol No-76, Issue 3, 1-10
International Conference on “Business Economics, Social Science & Humanities” BESSH-2015
ANALYSIS MODEL
Spatial Autocorrelation Test
To run the spatial regression analysis, the dependent variable
must be analyzed or draw a spatial autocorrelation. According to
the existing research results, home sales prices that are specified
in the dependent variable in this study, it is common to show the
spatial autocorrelation properties. To examine the data that are
used in this study whether show the same nature abovementioned
or not, analyze by drawing a map. As a result, it had been divided
high price areas and low price areas as a <Figure 5>.
Figure 4. Apartment price map of the research area
Therefore, the research area‟s housing value, it is determined that
there is a spatial dependency and would need more for the exact
analysis through Moran‟s I coefficient statistical analysis(Moran,
1950). The research data was tested for globally spatial clusters
through Moran‟s global clustering test to check whether spatial
autocorrelation exists. Global Moran‟s I can be found through
formula (1).
𝐼 =𝑛 𝑤𝑖 ,𝑗 𝑧𝑖𝑧𝑗
𝑛𝑗=1
𝑛𝑖=1
𝑆 𝑧𝑖2𝑛
𝑖=1
𝑤ℎ𝑒𝑟𝑒, 𝑆 = 𝑤𝑖 ,𝑗𝑛
𝑗=𝑖
𝑛
𝑖=1
(1)
Here, 𝑧𝑖 is the difference between the values of each feature and
the mean (𝑥𝑖 − 𝑋 ), while 𝑤𝑖 ,𝑗 represents the spatial dependency
between feature 𝑖 and 𝑗. 𝑛 Represents the number of all the
features. The index 𝐼 derived through this formula ranges from 0
to ±1 and the degree of spatially-correlation is stronger the closer
the index is to ±1 and away from 0 (Li et al, 2007).
In this formula, 𝑋 is a matrix of (𝑛 × 𝑘), 𝑌 is the (𝑛 × 1) vector
and 𝑊(𝑢𝑖 ,𝑣𝑖) is a matrix of (𝑛 × 𝑛). The diagonal elements in
𝑊(𝑢𝑖 ,𝑣𝑖) represent the geographical weights of centric
coordinate (𝑢𝑖 , 𝑣𝑖)‟s region 𝑖 and the adjacent 𝑛 observing points.
In GWR, 𝑊, the SWM, uses similar principles as the SWM used
to calculate Global Moran‟s I and Getis Ord G Index, but is
different as it is endogenously presumed.
Moran‟s I of the average sale price was calculated at 0.618 (Z-
score is 9.11 SD). Thus, it was possible to grasp that the high
level spatial dependence is shown.
Generating Spatial Weights Matrix(𝐖𝐮)
Taken together the results of the phenomenon analysis, not only
adjacency, but also other influences affect to the housing market
in Seoul. Therefore, It is not rational that calculate distance only
when generating the SWM for taking into account the spatial
effects. The new expression configured to solve these problems,
a <Formula 2> described below. From each nonnegative matrix
𝑊 = (𝑤𝑖𝑗 : 𝑖, 𝑗 = 1,2,… ,𝑛) is a possible SWM which
summarizes spatial relationships among 𝑛 spatial units.
𝑊𝑑 = (𝑤𝑖 ,𝑗 :𝑑𝑖𝑗−2)
(2)
Kabsung Kim /BESSH--2016/Full Paper Proceeding Vol No-76, Issue 3, 35-45
International Conference on “Business Economics, Social Science & Humanities” BESSH-2015
42
𝑊𝑟 = (𝑤𝑖 ,𝑗 : each element in < 𝑇𝑎𝑏𝑙𝑒 3 >)
𝑊𝑢 = 𝑊𝑑 +𝑊𝑟
Focus on the <Formula 2>, which is for composition matrix𝑊𝑢 ,
some variations were applied to the conventional method. First,
obtain the original distance weighting matrix was names𝑊𝑑 . After
that utilizing the correlation coefficient of <Table 3> to make
housing price relational matrix. 𝑊𝑟 is calculated from row
standardization applied housing price relational matrix.𝑊𝑢 is
finally derivate summing 𝑊𝑑 and𝑊𝑟 , and conducted a row
standardization again. Calculation formula of row standardization
is <Formula 3>.
𝑤𝑖 ,𝑗𝑠 =𝑤𝑖 ,𝑗 𝑤𝑖 ,𝑗𝑗
(2)
Here, excluded self-explanation by assign that 𝑤𝑖𝑖 = 0 for all
𝑖 = 1,2,… ,𝑛. So 𝑊 has a zero diagonal element. Each element of
the 𝑊𝑢 is closer the distance between the apartment, and the
higher the correlation of changes in the buying and selling price,
with a greater value. The value of the case of the reverse is
lowered.
Applying Spatial Lag Model
<Formula 4> is basically accept the existing spatial lag model
utilizes 𝑊𝑢 calculated in <Formula 2>. The reason for using
spatial lag model is, in the study of Kim (2000), because it has
been demonstrated that already a spatial lag model in the housing
market of Seoul have explanatory power is superior to the spatial
error model. Spatial lag model assumed that affected not only the
explanatory variable of the internal housing prices in this area, but
also the price of the house neighbored region(Anselin, 1988).
Y = 𝜌𝑊𝑢𝑌 + 𝑋𝛽 + ϵ
𝑊ℎ𝑒𝑟𝑒,
𝑌:𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒𝑠
𝑋: 𝑖𝑛𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒𝑠
𝑊𝑢 :𝑈𝑛𝑖𝑓𝑖𝑒𝑑 𝑆𝑊𝑀
𝜌,𝛽: 𝑝𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟𝑠 𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑒𝑑
𝜖: 𝑖. 𝑖.𝑑.
(4)
In this study, to adjust the ratio of the relative importance, from
(𝑑𝑖𝑠𝑡. 𝑖𝑚𝑝𝑜𝑟𝑡𝑎𝑛𝑐𝑒 ∶ 𝑐𝑜𝑟𝑟. 𝑖𝑚𝑝𝑜𝑟𝑡𝑎𝑛𝑐𝑒 = 0.2: 0.8)and
𝑐𝑜𝑟𝑟. 𝑖𝑚𝑝𝑜𝑟𝑡𝑎𝑛𝑐𝑒 𝑑𝑖𝑠𝑡. 𝑖𝑚𝑝𝑜𝑟𝑡𝑎𝑛𝑐𝑒 = 0.2: 0.8) be subjected to
analysis. At this time, the RMSE is to look for the best
combination of high explanatory power as finding the smallest.
The statistics to refer to in order to determine what method is
most effective in calculating the SWM is Akaike‟s Information
Criterion (AICc), and it is generally agreed that SWM with
minimal AICc is the best method (Fortheringham et al, 2002).
ANALYSIS RESULTS
Analysis, three models were performed, respectively. First is the
OLS analysis commonly used, Second, 𝑆𝐿𝑀𝑑 using distance
based SWM (𝑊𝑑 ), the third is a 𝑆𝐿𝑀𝑢 using Unified SWM (𝑊𝑢 ).
The first model has been added to clarify that to derive the results
of spatial regression was excellent much compared to the OLS.
The second model and the third model, by comparing both, were
analyzed separately in order to demonstrate the utility of 𝑊𝑢 .
Here, the weight of the correlation between the weights of the
physical distance, respectively 0.4:0.6. This decision was reason
in this way, will be described later.
Looking at the spatial regression analysis result of <Table 4>,
𝑆𝐿𝑀𝑢 shows the highest R-squared value and lower standard
error than the others (𝑂𝐿𝑆, 𝑆𝐿𝑀𝑑 ). Since the Durbin-Watson
numerical value has increased favorable direction, to 2, 𝑆𝐿𝑀𝑢 is
relatively free in multicollinearity problem. The spatial
coefficient is in 𝑆𝐿𝑀𝑑 displayed 0.384, in 𝑆𝐿𝑀𝑢 is 0.327. It
means that in 𝑆𝐿𝑀𝑢 , correlation effect became independent.
AICc values and Maximum likelihood based measure (ML)
represents the fit of the model 𝑆𝐿𝑀𝑢 is more suitable than 𝑆𝐿𝑀𝑑 .
Conventionally ML is larger, AICc is smaller means that the
model is appropriate.
As the individual variables point of view, whether the house's
location belongs to the U-City or not because it did not meet the
significance, it has been analyzed that does not have a significant
effect on the housing prices of up to now.
Another special variable, there is a need to look at the variables
age of the apartment. Because in OLS and 𝑆𝐿𝑀𝑑 analyses
demonstrated a positive sign, but in 𝑆𝐿𝑀𝑢 analysis, it was shown
a negative sign. Because the coefficient from 𝑆𝐿𝑀𝑢 does not
satisfy a significance level, so it is difficult to have a statistical
significance. However, it can put a meaning to the calibration
apartment variable in a common sense way while the process of
building the 𝑆𝐿𝑀𝑢 .
Kabsung Kim /BESSH--2016/Full Paper Proceeding Vol No-76, Issue 3, 35-45
International Conference on “Business Economics, Social Science & Humanities” BESSH-2015
43
In this study, first, look at the correlation of the sales price for the
entire apartment complex in Seoul, and it was examined whether
can be explained only by the existing autoregression model. In the
result of the analysis, apartment transaction price could have been
formed not the specific phrase proximity only, but the some types
of correlation between specific districts also. In the existing
adjacency based SWM, it cannot reflect these correlations.
Because there can be described the effect on the area outside the
range of influence of the distance-decay function. To overcome
these limitations, we proposed a new method for creating SWM.
Next, calculate the RMSE for each model to statistically compare
the explanation power of the model. The result is 1.24 𝑆𝐿𝑀𝑑 and
𝑆𝐿𝑀𝑢 is calculated as 0.95. So, 𝑆𝐿𝑀𝑑 has a higher predictive
power than 𝑆𝐿𝑀𝑢 .
Like Abovementioned, the results of the spatial regression
(<Table 4>) has stated that using specific ratio between physical
distance and correlation coefficient. This is determined after
going change the relative weights of the two forces that are
TABLE IV
THE RESULT OF SPATIAL REGRESSION ANALYSIS
Variables
Standardized Coeff.
𝑂𝐿𝑆 𝑆𝐿𝑀𝑑 𝑆𝐿𝑀𝑢
Spatial Coefficient - 0.384**
0.327**
Intercept -51,928.2**
-62,151.4**
-57,837.3**
Physical
Character.
Rebuild. is Scheduled 0.019**
0.022**
0.022**
Total Buildings 0.107**
0.122**
0.123**
Area for Excl. Use 0.799**
0.594**
0.587**
Num. of Rooms -0.044**
-0.026**
-0.018**
Num. of Baths 0.007**
0.001**
0.002**
Total Household 0.035**
0.029**
0.028**
Parking Lot/HH 0.023**
0.028**
0.026**
Porch 0.027**
0.024**
0.024**
Aging of Apts. 0.094**
0.010**
-0.006**
Environ.
Character.
Dist. to General Hosp. -0.026**
-0.029**
-0.029**
Dist. to Pri. School -0.010**
-0.015**
-0.014**
Dist. to High School -0.033**
-0.037**
-0.039**
Dist. to Subway St. -0.116**
-0.104**
-0.105**
U-City Available 0.052**
0.048**
0.045**
Diagnosis
Values
Standard Error 20,589 17,047 15,680
D-W 1.58 1.62 1.71
R2 0.72 0.74 0.78
ML -23,056 -19,159 -17,233
AICc 39,593 36,101 34,816
**, *: .01 and .05 significance level, respectively
Kabsung Kim /BESSH--2016/Full Paper Proceeding Vol No-76, Issue 3, 35-45
International Conference on “Business Economics, Social Science & Humanities” BESSH-2015
44
included in 𝑊𝑢 compares the fit of the model.
The standard error of the <table 5> is smallest when
(𝑑𝑖𝑠𝑡. 𝑖𝑚𝑝𝑜𝑟𝑡𝑎𝑛𝑐𝑒 ∶ 𝑐𝑜𝑟𝑟. 𝑖𝑚𝑝𝑜𝑟𝑡𝑎𝑛𝑐𝑒 = 0.4: 0.6) and r-
squared value is highest in the same condition. RMSE also
represents a similar flow. So we decided respectively 0.4 and 0.6
is the best ratio after several tests.
TABLE V
THE RESULT OF SPATIAL REGRESSION ANALYSIS
CONCLUSION
In this study, first, look at the correlation of the sales price for the
entire apartment complex in Seoul, and it was examined whether
can be explained only by the existing autoregression model. In the
result of the analysis, apartment transaction price could have been
formed not the specific phrase proximity only, but the some types
of correlation between specific districts also. In the existing
adjacency based SWM, it cannot reflect these correlations.
Because there can be described the effect on the area outside the
range of influence of the distance-decay function. To overcome
these limitations, we proposed a new method for creating SWM.
As a result of spatial regression analysis utilizing the new SWM
that reflects the price correlations, statistical significance was
higher relatively as compared with the Ordinary Least Square
(OLS) model. Further, it also had a smaller RMSE when using the
modified SWM, so it is verified the analysis results are
statistically valid. Thus, it is possible to conclude that real
explanatory power of SWM considering the correlation of price
fluctuations is stronger than existed simpler SWM such as
calculated from distance or adjacency only.
In addition, the best experimental result was derived from various
ratio attempt to change between physical distance significances
and relatively correlation significances. Result ratio is 0.4:0.6,
respectively. At those values, both fit and explanatory power were
the highest. Conversely, that the changes caused by other price-
related factors, such as Newtown construction or U-City
construction have accounted for 60 percent, it can be said quite
encouraging results.
However, it is limited in the fourth quarter of 2015 data used in
this study can point out the limitations that are close to almost
cross-section analysis. If it is possible to perform a future time
series analysis, will be able to SWM proposed in this study
receive a backing using a more robust model. Also, the analysis
did not reveal the source of the correlational effect can be pointed
out as well threshold. That is, another limitation of the analysis
was only revealed only how much correlated its influence and
there is missing detailed analysis.
ACKNOWLEDGMENT
This work is financially supported by Korea Minister of Ministry
of Land, Infrastructure and Transport (MOLIT) as 「 U-City
Master and Doctor Course Grant Program.
REFERENCES
[1] Anselin, L. (1988). Spatial econometrics: methods and models. Dordrecht, Kluwer Academic Publishers.
[2] Cho, M. H. (1998). The analysis of the effect of incineration facility on the land price of Peripheral area. Seoul National
University M. S. Dissertation.
[3] Choi, M. J., & Lim, Y. J. (2001). Empirical Analyses of the Relationships between Household Characteristics and Preference of
Residential Location and Housing Types. Journal of the Korea Planners Association, 36(6), 69-81.
[4] Choi, M. S., Kim, E. J. & Bang, J. W. (2003). Spatial dependence of apartment sale prices in seoul: an application of spatial
econometric model. Journal of the Korean Regional Science Association, 19(3), 61-80.
Class Goodness of Fit
Physical
Distance
Correlation
Coefficient
Standard
Error R
2 RMSE
0.2 0.8 15,871 0.75 0.6
0.3 0.7 15,795 0.76 0.6
0.4 0.6 15,680 0.78 0.6
0.5 0.5 15,734 0.76 0.5
0.6 0.4 15,861 0.76 0.5
0.7 0.3 16,029 0.74 0.4
0.8 0.2 16,253 0.72 0.2
Kabsung Kim /BESSH--2016/Full Paper Proceeding Vol No-76, Issue 3, 35-45
International Conference on “Business Economics, Social Science & Humanities” BESSH-2015
45
[5] Choi, Y., Son, T. M. & Kim, S. I. (2000). Effects of NIMBY Facilities on Residential Property Values : Prison & Power Transit
Station Cases. Journal of Korea Planners Association, 35(1), 151-163.
[6] Choi, Y. & Kwon, Y. H. (2004). The impact of educational environment on multi-family attached house prices using hierarchical
linear model. Journal of Korea Planners Association, 39(6), 71-82.
[7] Getis, A. & Aldstadt, J. (2004). Constructing the spatial weights matrix using a local statistic. Introduction to Geography 10th ed.
Dubuque, IA:McGaw-Hill.
[8] Gillen, K., Thibodeau, T., & Wachter, S. (2001). Anisotropic autocorrelation in house prices. The Journal of Real Estate Finance
and Economics, 23(1), 5-30.
[9] Hwang, Y. W. & Kang, G. C. (2009). The study on the influences of rental housing to busan housing market and its characteristics.
The Korea Spatial Planning Review, 61, 287-299.
[10] Aldstadt, J. & Getis, A. (2006). Using amoeba to create a spatial weights matrix and identify spatial clusters. Geographical
Analysis, 38, 327-343.
[11] Jung, S. Y. (2006). Impact of educational variable on apartment price in Seoul. Journal of Korea Planners Association, 41(2), 153-
166.
[12] Kim, J. W. 2000. Estimation of regression coefficient and verification of spatial autocorrelation in the housing market. Kyong Je
Hak Yon Gu, 48(2), 155-173.
[13] Kim, J. Y. (2003). Estimation of housing price function using hierarchy linear model. Journal of Korea Planners Association
38(7), 223-235.
[14] Kim, J. Y. and Yoon, D. K. (2004). Comparison with the estimation method of housing price functions. Korean Appraisal Review,
4(1), 207-227.
[15] Lancaster, K. (1966). A New approach to consumer theory. Journal of Political Economy, 74, 132-157.
[16] Lee, K. H., Seo. S. J. (2009). A Study on the Regional Disparity between the Major 3 Gangnam Districts and the Major 3 Gangbuk
Districts in Seoul. Korean Public Management Review, 23(4), 357-381.
[17] Lee, S. H. & Jeon, K. K. (2012). A study of the influences of urban infrastructure on housing prices: a hierarchical linear model.
Journal of Korea Planners Association, 47(4), 193-204.
[18] Lee, C. R, Park, K. H. (2013). The effects of neighborhood segmentation on the adequacy of a spatial regression model. Journal of
Korean Geographical Society, 48(6), 978–993.
[19] Li, H., Calder, C. A., & Cressie, N. (2007). Beyond Moran's I: testing for spatial dependence based on the spatial autoregressive
model. Geographical Analysis, 39(4), 357-375.
[20] MOLIT (Ministry of Land, Infrastructure and Transportation). (2013). Summarize of the U-City test-bed projects. Press Material,
1-5.
[21] Moon, J. H., Seo, K. C. & Lee, S. H. (2006). A study of the influence of public housing on the land price of community. Journal of
Korea Planners Association, 41(4), 103-116.
[22] Moran, P. A. P. (1950). Notes on continuous stochastic phenomena. Biometrika, 37(1), 17–23.
[23] Oh, Y. K. Kang, J. K. & Kim, J. M. (2014). Analysis of regional characteristics that affect housing prices using a gwr model -
focused on busan metropolitan city. The Journal of Korea Tax Accounting Research, 40, 1-17.
[24] Park, H.S., Ahn, J. A. (2005). “A study on the spatial effects of apartment prices using spatial dependence. Journal of KREAA
11(2), 119-128.
[25] Rosen, S. (1974). Hedonic prices and implicit markets: product differentiation in pure competition. Journal of political
economy, 82(1), 34-55.
[26] Woo, A. Y. (2005). Effects of social mix with public rental housing on housing prices. Seoul National University M. S.
Dissertation.
[27] Yoon, H. M. & Jung, S. Y. (2013). The effects regional characteristics of housing environment in seoul upon housing prices.
Journal of the Korea Real Estate Analysts Association, 19(4): 235-253.